Mike Winter built a full-AI combat robot, put it in the arena against his daughter — a BattleBots driver for decades — and watched it learn not to expose the corner that kept getting hit.
From there we talk factories and marketing that would be crazy not to use AI, the diorama work on his desk, and the idea that language models are only one kind of model. Mike says he doesn’t think in a language — and I push the case for non-language creative models alongside the LLM stack.
Transcript
Mike: I think it will be interesting because I did a contest called AI or Die. And it was about a human versus robot, and it was just a few years ago, and I had an AI robot made for my Battle Bots experience. It weighs 15 pounds, but it had a dangerous weapon on it, full AI, some cameras, but it was my own kind of AI.
If something attacks me, remember what happened, and don’t let that happen again. Anything that moves: kill, right? I had my daughter, who’s driven Battle Bots for 20 years. We got a big arena in a fancy place. I put my robot in there. I walked away, sat in the chair in the audience. My daughter went there. They had a battle.
Ukraine should have seen this battle, because it would have helped what they were doing. The robots attacked each other. The AI robot had realized what my daughter was doing. My daughter realized where the cameras were. The robot couldn’t see it, and the robot realized, ‘Every time I exposed this corner to the other robot, I’m getting hit.’ After one minute, the AI was so much more powerful. And it was doing things 10,000 to a second. Decisions were being made.
Pablos: That’s what’s possible. I think there’s so many industrial-grade versions of that underway now. Amazing what you could do with it.
Mike: If I had a factory, too, I suspect that you don’t know a lot about what’s happening with AI, because it’s happening so fast right now. But if you have any kind of production, or especially marketing, you’d be crazy not to be using AI.
Pablos: I think there’s vanishingly few things left where that is true. It’s helping with everything now. I’m excited about that. I think there’s this crazy sort of discourse around it, but if you just think about how AI is the most democratized technology of all time. It can literally explain to a five-year-old how it works, how to use it, what do you need at your level, in your language. It can do that for anyone.
And it’s somewhere between free and 20 bucks a month for anyone on Earth who has found their way online.
Mike: I feel so naive about what AI can do. I consider myself a pretty well-read person, using AI because it’s a blast. I am constantly, every moment, caught off guard. I’m making a diorama. I asked him some questions, like, ‘What color paint should I use on this figure?’ It’s like 10 steps ahead of me.
What’s going to be happening in a week or two is, I will come into my room and it will be running Fusion 360, Blender, or something like that, and it will be telling my 3D printer to make things. And then it will have me — I’ll be the grunt: ‘Take this out of the 3D printer.’
Pablos: Chat GPT doesn’t work this way very well, but some of the other so-called harnesses are very proactive. They try to figure out what you would do next if you weren’t sleeping, and then they go do it. Their AI is tired of waiting for you to wake up.
It figures out, ‘Probably the next thing to do would be to model it. And then probably the next thing to do would be to print it. And it looks like there’s a printer over here. I’ll do that. And that way, when you wake up, it’s ready to go.’
I’m using the one called Hermes, which you’ve probably heard of. And it runs 24 seven. It constantly looks for things to do. Every morning I wake up, I look and see. It’s, ‘I thought I would build this for you and see what you think.’ ‘Yeah, that sucks. Throw it out.’
Mike: A lot of people are doing workflow. I use it as a kind of a search engine for a search engine.
Pablos: It wants to be a colleague at this point. It’s got a PhD in diorama, so it thinks it ought to be put to good use. Hard to find one these days, but you found one.
Mike: In France, writing on the cave wall, that’s considered writing. The pictures of people chasing animals or maps or whatever, that’s not considered artwork. It’s considered proto-writing. But people carried amulets, and back then they had little bison or mastodons or whatever they were chasing. They made reenactments in the cave floor.
Pablos: A reenactment with little figurines would be way more useful than writing in that context, probably.
Mike: Also, the most common artwork made in the world by far are dioramas. But once you start thinking about it, you’ll see that: every building made, every planning, every war.
Pablos: It always starts out as a product you’re going to make. There’s a styrofoam iPhone in Cupertino somewhere. What utility is there in making dioramas for you, or is this just a creative endeavor?
Mike: I started making them when I was very young. I don’t think in a language. Our consciousness — we think in terms of language: ‘I have to go to the post office, and then I have to pick up the lasagna before I go over to this person’s house — do A, B, C, B, and E.’ But a lot of times when you’re making stuff, you should not be using that, because you’re using patterns that maybe others give you. It gets you out of the thinking mode that can be polluted by other people’s writing.
And it can really do that. I’ve been really lucky to get these jobs. When somebody would come into a room and say, ‘We gotta figure out how to make a phone that understands 3D so that people can figure out how to navigate a room or something like that,’ I would have my subconscious work. I’d blurt out things: ‘We should look at every possible picture of a room like that and know the dimensions of everything.’
I reckon if you have a company and you’re trying to do creative stuff, task your engineers with not thinking in the language, and make them come up with solutions unreasonably fast. ‘We’re gonna design the new Tesla, but it’s going to use this other kind of battery more efficiently. It also has to be a bigger part of your life. You have one minute to write out the plan.’
Pablos: To me, the whole exercise of making a diorama — it makes sense: early primitive form of art, maybe, or early art, because the job is to create a model. Diorama is a model.
What we have in our brains is a model that we’ve trained. Yours is different than mine. You trained on different stuff, but we each have our own model we’ve trained. And what we’ve always been trying to do with our computers is make a model. A spreadsheet is a model; it’s very limited. And video games are models. Fusion 360 is making models, and the LLMs are a model of language.
And language is a fungible means of describing different models, but it is not a functional way to make all kinds of models. Not a good way to make a model of the Tesla or Rocket or whatever. And so I think that what we’re held back by right now in AI is such an extreme emphasis on the language model, because it is generally applicable in the way that language is generally applicable. But no amount of language can replace a photograph or artwork. And that’s what people are missing right now. They need to make other kinds of models in order to make AI go other places.
Mike: I love that.